Supervision of Video Game Car Steering Implementing HORCNN Network

This research aims to make an algorithm that can autonomously drive a car in the perspective of a 3^rd person in a video game using the Convolutional Neural Network (CNN) approach. The Hybrid Object Recognition CNN (HORCNN) model has been programmed, combining two different networks. The environment and its parameters are extracted from an 800*600 windowed mode of the game. One frame is sent to a Neural Network Pipeline, first consisting of image processing, then feeding it into a Conv Net, the infamous Alex-Net. The output of that would be a vector containing probabilities of which direction to steer and whether to go forward or not. Parallelly, the same image is fed into the singleshot detector (SSD) network to get the objects detected in that particular video frame, and this information is used to label objects in that frame. The proposed model gets the absolute accuracy at the speed of 113 km/h, which can also simulate an actual autonomous vehicle.